Machine Learning Assisted Design for Active Cathode Materials

The traditional way of designing materials, including experimental measurement and computational simulation, are not efficient. Machine learning is considered a promising solution for material design in the recent years. By observing from previous data, machine learning finds patterns, learns from the patterns and predict the material properties. In this study, machine learning methods are used for discovering new cathode with better properties, includes crystal system learning and the property prediction. K-Folder cross-validation is used for finding the best training data with a limited dataset, nevertheless increasing the percentage of training data would ultimately result in better performance on prediction. It is found that, random forest gives the highest average accuracy in crystal system classification, meanwhile, extra randomized tree algorithm provides a higher averaged coefficient of determination and lower mean squared error in the regression model predicting electrical properties of cathodes. The random forest algorithm is chosen from a wide range of machine learning algorithms with the implementation of Monte Carlo validation. Based on the feature importance evaluation, oxygen contents are found to have the highest effects in determining capacity gravity and volume change in properties prediction.